Abstract

Prevailing wisdom asserts that one cannot rely on the cloud for criticalreal-time control systems like self-driving cars. We argue that we can, andmust. Following the trends of increasing model sizes, improvements in hardware,and evolving mobile networks, we identify an opportunity to offload parts oftime-sensitive and latency-critical compute to the cloud. Doing so requirescarefully allocating bandwidth to meet strict latency SLOs, while maximizingbenefit to the car.

 

Quick Read (beta)